Caching for Analytic Computations
Humans repeat stuff. Caching helps.
Normal caching policies like LRU aren't well suited for analytic computations where both the cost of recomputation and the cost of storage routinely vary by one million or more. Consider the following computations
# Want this
np.std(x) # tiny result, costly to recompute
# Don't want this
np.transpose(x) # huge result, cheap to recompute
Cachey tries to hold on to values that have the following characteristics
- Expensive to recompute (in seconds)
- Cheap to store (in bytes)
- Frequently used
- Recenty used
It accomplishes this by adding the following to each items score on each access
score += compute_time / num_bytes * (1 + eps) ** tick_time
For some small value of epsilon (which determines the memory halflife.) This has units of inverse bandwidth, has exponential decay of old results and roughly linear amplification of repeated results.
Example
>>> from cachey import Cache
>>> c = Cache(1e9, 1) # 1 GB, cut off anything with cost 1 or less
>>> c.put('x', 'some value', cost=3)
>>> c.put('y', 'other value', cost=2)
>>> c.get('x')
'some value'
This also has a memoize method
>>> memo_f = c.memoize(f)
Status
Cachey is new and not robust.